During development testing, Taklif.AI guardrails rejected roughly 12% of interests, 8% of input documents, and 3% of generated assignments¶

Claim · All claims
Evidence · 1 study · design r2 · q1
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q1 i? Across approximately 200 development-phase test inputs, interest guardrails rejected about 12% of submitted interests (including prompt injection attempts), assignment guardrails rejected about 8% of input documents, and output guardrails flagged about 3% of generated assignments. → Zaki Kurdya 2026

Evidence¶

Zaki Kurdya 2026¶

Zaki Kurdya, Mohammed Zuqlam, Salem Amassi, Shady Telbany, Motaz Saad. (2026). Taklif.AI: LLM-Powered Platform for Interest-Based Personalized College Assignments. https://arxiv.org/abs/2605.05842 doi:10.48550/arxiv.2605.05842

q1 · i? · design · r2

Development-phase guardrails testing with approximately 200 test inputs. Interest guardrails rejected "approximately 12% of submitted interests", assignment guardrails about 8% of input documents (e.g., lecture notes, syllabi), and output guardrails flagged about 3% of generated assignments. No effect size is printed.

"Theinterest guardrailsrejected approximately 12% of submitted interests, including single-character inputs, profanity, non-interest terms (e.g., “nothing,” “I don’t know”), and attempted prompt injection attacks"

Discussion¶

Related Claims¶

• Most student project reports exceed the BART model's 1024-token input limit, motivating a select-then-generate design because truncation can lose critical information — related
• In preliminary user acceptance testing, 84% of participants rated Taklif.AI's personalization feature as beneficial — related
• A guardrailed GPT-4 tutor with teacher-designed hints improves assisted practice performance more than a standard GPT-4 interface — related